PaperlayerView exact revision on arXiv

Source-linked AI summary

Estimating the Effects of Continuous-valued Interventions using Generative Adversarial Networks

Ioana Bica, James Jordon, Mihaela van der Schaar

arXiv:2002.12326v2cs.LGstat.ML

TL;DR

Estimating intervention effects from observational data is well developed for discrete treatments but less studied for continuous-valued interventions, where counterfactual outcomes are unobserved and selection-bias adjustment is more complex. SCIGAN modifies GANs to generate counterfactual outcomes and uses a hierarchical discriminator; semi-synthetic experiments show it outperforms existing benchmarks.

  • Problem

    Continuous-valued interventions are less studied than discrete treatments, while observational data provide only factual outcomes and make selection-bias adjustment more complex.

    Source: §1 · pp. 1–2

  • Method

    SCIGAN modifies GANs to generate counterfactual outcomes, using a hierarchical discriminator that separates factual-intervention and dosage decisions.

    Source: §1 · p. 2, §4.2 · p. 4

  • Results

    SCIGAN outperforms existing benchmarks in semi-synthetic experiments.

    Source: §1 · p. 2

  • Takeaways & Limitations

    SCIGAN provides a framework for estimating outcomes across continuous and many-level-discrete interventions.

    Source: §1 · p. 2

    The authors say significant testing, potentially through clinical trials, is needed before the method can be used in practice.

    Source: Broader Impact · p. 9

Abstract

from arXiv · show

While much attention has been given to the problem of estimating the effect of discrete interventions from observational data, relatively little work has been done in the setting of continuous-valued interventions, such as treatments associated with a dosage parameter. In this paper, we tackle this problem by building on a modification of the generative adversarial networks (GANs) framework. Our model, SCIGAN, is flexible and capable of simultaneously estimating counterfactual outcomes for several different continuous interventions. The key idea is to use a significantly modified GAN model to learn to generate counterfactual outcomes, which can then be used to learn an inference model, using standard supervised methods, capable of estimating these counterfactuals for a new sample. To address the challenges presented by shifting to continuous interventions, we propose a novel architecture for our discriminator - we build a hierarchical discriminator that leverages the structure of the continuous intervention setting. Moreover, we provide theoretical results to support our use of the GAN framework and of the hierarchical discriminator. In the experiments section, we introduce a new semi-synthetic data simulation for use in the continuous intervention setting and demonstrate improvements over the existing benchmark models.

1 Introduction

Estimating continuous interventions from observational data is difficult because outcomes for unobserved interventions are counterfactual and treatment assignment is non-random. SCIGAN modifies GANs to generate counterfactual outcomes, using a hierarchical discriminator and a generator for multiple interventions.

1 Introduction

  • Continuous interventions create an effectively unbounded set of counterfactuals, while non-random assignment makes selection-bias adjustment harder than for discrete treatments.Methods for adjusting selection bias in discrete settings cannot be easily extended to continuous interventions.Source: §1 · p. 2
  • SCIGAN adapts GANITE’s discriminator objective, distinguishing the factual outcome from generated counterfactual outcomes rather than classifying whole samples as real or fake.The paper further changes the generator and discriminator for continuous interventions.Source: §1 · p. 2
  • SCIGAN’s hierarchical discriminator separately identifies the factual intervention and dosage, while its multi-task generator accepts continuous parameters to learn heterogeneous response curves.The discriminator evaluates a finite set of points from each generated response curve rather than entire curves.Source: §1 · p. 2
  • SCIGAN outperforms existing benchmarks in semi-synthetic experiments and estimates outcomes for continuous and many-level-discrete interventions.The paper also introduces a semi-synthetic simulation for continuous interventions.Source: §1 · p. 2

2 Related work

Prior approaches to continuous-dose effects use generalized propensity scores or balancing methods, including neural networks with treatment and dosage-interval heads. SCIGAN uses GANs to generate counterfactuals and reports outperforming GPS and DRNets.

2 Related work

  • SCIGAN uses GANs to generate counterfactuals, unlike GPS and DRNets’ propensity-score or supervised bias-adjustment approaches, and experimentally outperforms both.DRNets use shared layers, treatment-specific layers, and dosage-subinterval heads, but their intervals are not determined dynamically.Source: §2 · p. 3
  • For treatment-response methods without a dosage parameter, dosage cannot simply be added as an input because its assignment is biased; the paper also relates this work to continuous-treatment policy optimization.Source: §2 · p. 3

3 Problem formulation

The formulation treats each intervention as a treatment-dosage pair and targets individualized potential outcomes for given covariates. Identifying those outcomes from observational data requires unconfoundedness and overlap.

3 Problem formulation

  • Each observation includes pre-treatment covariates and a factual treatment-dosage pair, with treatment type drawn from a discrete set and dosage from its treatment-specific space.The factual treatment is T_f = (W_f, D_f).Source: §3 · p. 3
  • For every treatment-dosage pair, the framework defines a potential outcome; the observed outcome is the one matching the received pair, and the others are counterfactuals.Source: §3 · p. 3
  • The target is an individualized dose-response function giving unbiased estimates of potential outcomes for each covariate vector and treatment-dosage pair.The paper denotes this function by µ(·).Source: §3 · p. 3
  • Identification assumes unconfoundedness, so treatment assignment is conditionally independent of potential outcomes given covariates, and overlap, so each treatment-dosage pair has positive assignment probability below one.Source: §3 · p. 3

4 SCIGAN

SCIGAN uses a modified GAN to generate counterfactual dose-response curves, then trains an inference network to estimate outcomes for new samples. Its hierarchical discriminator separates treatment and dosage judgments, and the paper gives a distributional guarantee for the learned counterfactuals.

4.1 Counterfactual Generator

  • The generator maps features, factual treatment and outcome, and noise to dose-response curves, with a supervised loss regularizing its factual-pair output.The curves provide generated outcomes for treatment-dosage pairs, including those not observed.Source: §4.1 · p. 4

4.2 Counterfactual Discriminator

  • The discriminator uses sampled points from generated curves and separates factual-treatment classification from dosage classification within each treatment.For the factual treatment, the sampled dosage set includes the observed dosage alongside random dosage levels.Source: §4.2 · p. 4, Figure 2 · p. 5

4.3 Inference Network

  • After learning the generator, SCIGAN trains an inference network on generated curves and original data to estimate outcomes for new samples.The inference network takes features and a treatment-dosage pair as inputs.Source: §4.3 · p. 5

4.4 Theoretical Analysis

  • At the game's global minimum, generated counterfactuals for unassigned treatment-dosage pairs have the true marginal outcome distribution conditional on features.The theorem concerns marginal distributions, not a guarantee about each individual's exact counterfactual.Source: §4.4 · p. 5

5 Architecture

The architecture combines a generator with hierarchical discriminators designed for sets of dosage-outcome points. The inference network shares the generator architecture but omits its factual-treatment, factual-dosage, factual-outcome, and noise inputs.

5 Architecture

  • The inference network uses the generator architecture but excludes factual treatment, factual dosage, factual outcome, and noise as inputs.Source: §5 · p. 6

5.1 Generator Architecture

  • A shared generator representation feeds treatment-specific heads that combine the representation with dosage to produce an outcome.The generator and its treatment heads are fully connected networks.Source: §5.1 · p. 6

5.2 Hierarchical Discriminator Architectures

  • The hierarchical discriminator uses permutation-invariant treatment-level processing and permutation-equivariant dosage discriminators operating on each treatment's dosage-outcome set.The treatment discriminator combines its set representation with features to output probabilities over treatments.Source: §5.2 · pp. 6–7

6 Evaluation

On semi-synthetic benchmarks, SCIGAN significantly outperforms every benchmark across all three datasets, maintains performance as discrete dosages increase, and outperforms benchmarks across treatment- and dosage-bias settings.

6.1 Experimental setup

  • Because counterfactuals are unobserved, the evaluation uses semi-synthetic outcomes generated from real-dataset features, three treatments with dose-response curves, and added Gaussian noise.The feature sources include TCGA, News, and MIMIC III; Table 1 describes the response curves and sets C = 10.Source: §6 · p. 7, §6.1 · p. 7, Table 1 · p. 7
  • The setup varies dosage-selection bias with α and treatment-selection bias with κ, then compares GPS, DRNet variants, and MLP baselines using MISE, DPE, and PE.Unless otherwise specified, κ = 2 and α = 2.Source: §6.1 · p. 7

6.2 Source of gain

  • Adding each component improves performance, and the full model performs best across both datasets and all metrics in the source-of-gain analysis.Table 2 presents the source-of-gain analysis on TCGA.Source: §6.2 · p. 8, Table 2 · p. 8

6.3 Benchmarks comparison

  • SCIGAN significantly improves over every benchmark across all three semi-synthetic datasets; for MIMIC, the comparison uses treatments 2 and 3 because of limited sample size.Table 3 reports individualized treatment-dose response performance on three datasets, with mean ± standard deviation.Source: Table 3 · p. 8, §6.3 · p. 8

6.4 Discrete dosages

  • As discrete dosage levels rise from 3 to 30, GANITE cannot handle more than 7, while SCIGAN maintains performance.The authors attribute SCIGAN’s performance to its hierarchical discriminator and invariant and equivariant layers.Source: §6.4 · p. 8, Figure 4 · p. 8

6.5 Treatment and dosage selection bias

  • On TCGA, SCIGAN consistently and significantly outperforms the benchmarks in PE across κ from 0 (no bias) to 10 (strong bias) and α from 1 (no bias) to 8 (strong bias).Other metrics are in Appendix M.8.Source: §6.5 · p. 9

7 Conclusion

The conclusion presents SCIGAN as a framework for estimating response curves for continuous interventions from observational data, while noting its sample-size requirement and directions for future work.

7 Conclusion

  • SCIGAN estimates response curves for continuous interventions from observational data using a modified GAN framework with a hierarchical discriminator.The paper provides theoretical justification for both the modified GAN framework and the hierarchical discriminator.Source: §7 · p. 9
  • SCIGAN requires at least a few thousand training samples, a limitation the authors associate with neural networks and GANs.Source: §7 · p. 9
  • Future work could make the invariant and equivariant layers more expressive and extend the approach to temporal causal inference with sequences of continuous interventions.Source: §7 · p. 9

Broader Impact

The paper frames continuous-intervention estimation as a practical but difficult extension of causal inference and distinguishes individual counterfactual estimation from policy evaluation. It presents SCIGAN’s design alongside its theoretical scope and the need for further clinical testing before practice.

Broader Impact

  • The paper calls for substantial testing, potentially including clinical trials, before its method is used in practice, and recommends clinician support rather than replacement.It notes that incorrect estimates could have life-threatening implications and that intervention-effect estimates could also be misused to select harmful outcomes.Source: Broader Impact · p. 9

A Expanded Related Works

  • Prior continuous-dose methods include generalized propensity scores, balancing methods, and DRNets, while one cited approach uses interventional data that SCIGAN does not assume.The related work also distinguishes methods for discrete treatments and notes that separate regressions do not extend to an infinite treatment space.Source: A Expanded Related Works · p. 13

A.1 Comparison with GANITE

  • SCIGAN samples finite sets of intervention-outcome points for its discriminator and uses a permutation-invariant set representation, avoiding whole response curves or discretizing the intervention space.Its hierarchical discriminator separately predicts the factual intervention and dosage; its multitask generator models treatment-specific response curves.Source: §A.1 · p. 13

A.2 Off-policy evaluation and policy optimization with continuous treatments

  • Off-policy evaluation estimates policy value, whereas individual treatment-effect estimation targets each sample’s outcomes under every possible action and can support decisions beyond policy evaluation.The paper says SCIGAN assumes no prior form for the response curve, unlike a cited doubly robust estimator that depends on such knowledge.Source: §A.2 · pp. 13–14

B Notation

  • The notation defines the factual intervention as a treatment-dosage pair and the counterfactual outcome function as mapping interventions to outcomes.It also names the generator, discriminator, and inference network used in the paper’s framework.Source: B Notation · p. 14

C Proofs of Theoretical Results

The theoretical results characterize what the discriminator games recover and when the hierarchical game shares an optimum with the single-discriminator game. The accompanying procedures train the generator and inference network, while the appendices specify permutation-aware and single-discriminator designs.

C Proofs of Theoretical Results

  • At an optimum, generated counterfactuals for unobserved treatment-dosage pairs have the true conditional marginal outcome distribution; the hierarchical-game equivalence additionally requires conditional independence across treatment response curves.The single-discriminator result is sufficient for estimating the mean outcome because it concerns the marginal outcome distribution.Source: C Proofs of Theoretical Results · pp. 15–16

D Counterfactual Generator Pseudo-code

  • Generator training samples factual data, noise, and dosage sets, then updates dosage and treatment discriminators alongside the generator.The supplied algorithm fragments show separate dosage-discriminator and treatment-discriminator updates before generator updates.Source: D Counterfactual Generator Pseudo-code · p. 17

E Inference Network

  • The inference network maps features and interventions to outcomes and is trained on factual outcomes or generator-produced counterfactuals, depending on whether the intervention is factual.Training samples random dosage sets and uses the learned generator to supply counterfactual targets.Source: E Inference Network · p. 18

F.1 Definitions of Permutation Invariance and Permutation Equivariance

  • Permutation invariance makes a function’s output unchanged by reordering a set input, while permutation equivariance reorders its outputs correspondingly.The text also states that composing with an invariant function preserves invariance and composing two equivariant functions preserves equivariance.Source: §F.1 · p. 19

G Single Discriminator Model

  • The single discriminator is trained to output the probability that each treatment-dosage pair is factual, conditional on features and generated intervention-outcome inputs.Its loss depends on the generated outcomes, and the generator minimizes its GAN loss together with a supervised loss.Source: §G.1 · p. 20

G.1 Single Discriminator

  • The single-discriminator objective trains one network to identify the factual treatment-dosage pair among the generated candidates.The generator’s objective combines its GAN loss with the supervised loss, using the optimal discriminator.Source: §G.1 · p. 20

G.2 Single Discriminator Architecture

The single-discriminator architecture combines permutation-invariant context across treatment-specific dosage sets with permutation-equivariant outputs for each treatment’s dosage candidates.

G.2 Single Discriminator Architecture

  • The architecture first forms an invariant representation from treatment-specific dosage-outcome sets, then uses treatment-specific heads to produce equivariant outputs for each set.Each head uses the features and invariant representation as auxiliary input; Figure 6 depicts the single-discriminator architecture.Source: §G.2 · p. 20, Figure 6 · p. 20

H Ablation Studies architectures

  • The ablation architecture removes multitask generator components and the discriminator’s invariant and equivariant layers.Figure 7 depicts these ablation-study architectures.Source: Figure 7 · p. 21

I Dataset descriptions

The experiments describe three observational datasets, dosage-assignment bias, benchmarks, and evaluation metrics, followed by analyses of model gains and dosage-sample sensitivity.

I Dataset descriptions

  • Across News, TCGA, and MIMIC III, the study uses news-item features, cancer-patient gene expression, and ICU covariates, respectively, to represent continuous intervention settings.The datasets contain 10,000 News items, 9,659 TCGA samples, and 3,000 MIMIC III patients; the splits are 64/16/20% for training, validation, and testing.Source: I Dataset descriptions · p. 21

J Dosage bias

  • Dosage assignment uses a Beta distribution whose concentration parameter controls selection bias; Beta(1, 1) gives uniform, feature-independent dosages and no selection bias.As α increases, dosage variance decreases and sampled dosages lie closer to the optimal dosage.Source: J Dosage bias · p. 22

K Benchmarks

  • Benchmarks include DRNet, GPS, and MLP variants, with validation-based hyperparameter optimization and matched network architectures for the MLP comparisons.SCIGAN uses 5,000 GAN iterations and 10,000 inference-network iterations, selected to ensure convergence of the stated losses.Source: K Benchmarks · p. 23

L Metrics

  • The evaluation uses MISE for outcome accuracy across dosage space, DPE for optimal-dosage identification, and PE for selecting the optimal treatment-dosage pair.Metrics are computed on a held-out test set.Source: L Metrics · p. 25

M.1 News results for source of gain

  • The News source-of-gain analysis reports metrics as mean ± standard deviation, but the supplied passage gives no comparative values.Table 8 is identified as the source-of-gain analysis for News.Source: Table 8 · p. 26

M.2 Investigating hyperparameter sensitivity (nw)

  • On TCGA, the single discriminator drops across metrics around nw = 9, when its output dimension is 27; the hierarchical discriminator remains more stable through nw = 19, while News also shows a marked single-discriminator decline at high nw.Figures 9 and 8 respectively present the TCGA and News comparisons as the number of dosage samples increases.Source: Figure 9 · p. 26, Figure 8 · p. 26, §M.2 · p. 26

M.3 Dosage Policy Error for Benchmark Comparison

The appendix reports supplemental evaluations spanning dosage-policy error, treatment-count and sample-size variation, discrete-dosage comparisons, mixed treatment types, and selection bias.

M.3 Dosage Policy Error for Benchmark Comparison

  • Table 9 reports Dosage Policy Error for individualized treatment-dose response across three datasets, marking the best method per dataset and statistically significant improvements.The passage identifies these DPE results as corresponding to Section 6.3 of the main paper.Source: Table 9 · p. 27, §M.3 · p. 27

M.4 Varying the number of treatments

  • Adding three or six treatments, the experiment reports MISE, DPE, and PE on TCGA-6 and TCGA-9; SCIGAN training uses three dosage samples.Table 10 covers performance of SCIGAN and benchmarks at six and nine treatments.Source: §M.4 · p. 27, Table 10 · p. 27

M.5 Sample efficiency

  • The MIMIC sample-efficiency analysis compares SCIGAN trained with different sample counts on a fixed test set, reporting metrics as mean ± standard deviation.The test set is the same across training-sample conditions.Source: §M.5 · p. 27, Table 11 · p. 27

M.6 Discrete dosage set-up and comparison with GANITE

  • For <6 dosages, SCIGAN performs similarly to GANITE; above 6, it significantly outperforms GANITE, while SCIGAN improves then stabilizes around 12 dosages.The authors attribute GANITE’s degradation to its single discriminator, and SCIGAN’s weaker performance at low dosage counts to excess complexity and unnecessary dosage subsampling.Source: Figure 10 · p. 28, §M.6 · p. 28

M.7 Mixing dosage and no-dosage treatment options

  • SCIGAN extends to mixed dosage and no-dosage options via treatment-discriminator routing and generator heads without dosage inputs; Table 12 reports performance in this setting.The authors state that SCIGAN handles this setting and lends itself naturally to potentially mixed options.Source: §M.7 · p. 29, Table 12 · p. 29

M.8 Additional results on selection bias

Related papers

Loading 2002.12326v2…